1 Research outline

This notebook is associated with Article A, the main points of which are outlined here. All the result figures of the article in question can be generated using the code present in this notebook. Please refer to the article itself for more details about the methods.

The aim of the study is to investigate the mechanisms of resistance to BRAF inhibition in BRAF-mutated cancers. In particular, we are interested in colorectal cancer and melanoma which show different sensitivities despite some molecular similarities.

Graphical abstract of the complete study

Graphical abstract of the complete study

For this purpose, a generic logical model summarising the main signalling pathways around BRAF was constructed from the literature. This model is then personalized using cell line data (from Cell Model Passports). This results in as many personalized mechanistic models as there are cell lines. Simulations are then performed with these personalized models to see how they respond to BRAF node inhibition. This document is at the end of this pipeline since it will import the results of these simulations and compare them to the sensitivities observed experimentally.

2 Import simulations results and screening data

We first import simulation results from personalized models and drug/CRISPR screening files.

## [1] "All imports OK"

3 Clinical characterization of cell lines

We retrieve data from different screenings and with different metrics. Is there any consistence between all these values? We plot the correlation for the different measures across datasets and metrics. We focus on BRAF, TP53 and PIK3CA, the three targets mentioned in the article

Indeed, we observe some correlation clusters. When there are different drugs inhibiting the same target, their sensitivities are highly correlated. Besides, drug and CRISPR sensitivity outcomes are anti-correlated, especially for BRAF. This makes sense because of the definition of the metrics involved (sensitive cell lines have low AUC in drug screening and high scaled Bayesian factor in CRISPR).

4 Data reprocessing

Now we re-process the data to ease following analyses. Since the model contains a read-out node named Proliferation we will use it as a proxy to validate with experimental sensitivities.

We also define a normalised variable based on Proliferation level without any drug inhibition (i.e \(Proliferation_{normalised} = Proliferation_{withDrug} / Proliferation_{withoutDrug}\))

5 Simulations: a first quantitative approach with correlations

We first compare the Proliferation outcome from personalized models with the experimental sensitivities to BRAF inhibition.

We use different Proliferation proxies: the one from models without drug (resp. CRISPR) inhibition, the one from models with drug (resp. CRISPR) inhibition, and the normalised one. We also compare different personalization methods: with mutations only, with RNA only, and with mutations and RNA together.

5.1 All correlations for BRAF

On the following plot correlation coefficients are shown only for significant correlations. B panels correspond to Drugs screening and C panels to CRISPR screenings.

Some interesting points:

  • We observe very significant correlations for both drug and CRISPR screening
  • Normalised Proliferation appeared as the best proxy for drug sensitivity
  • RNA alone is usually not significant but its conjunction with mutation usually results in better results

5.2 Correaltions for other targets

Here are additional plots for other targets;

Since the present model has been designed around BRAF, its regulators have been carefully selected and implemented, which is not necessarily the case for other nodes of the model. Therefore, these personalized models can be used to study how comprehensive the descriptions of the regulation of other nodes or parts of the model are. Thus, model simulations show that response trends to TP53 inhibition are consistently recovered by the model but the simple regulation of p53 in the model results in coarse-grained patterns, although slightly improved by addition of RNA data. Similar analyses regarding the targeting of PIK3CA (in CRISPR data) simulated, in the model, by the inhibition of PI3K node, can be performed. Low correlations are an indication highlighting the insufficient regulation of the node.

5.3 Summary and publication plot

For drugs we will focus on one drug only which is PLX-4720 (very similar to Dabrafenib in many aspects, no particular criterion to distinguish) and AUC metric (less sensitive to extrapolation). For CRISPR screening we will focus on CC2 dataset, more balanced in CM and CRC. For output we will also focus on normalised Proliferation scores:

Here is the pruned version of the plot for publication:

6 Explore the results with scatter plots

Now we would like to have a look at more precise patters beyond correlation coefficients.

6.1 Simple plots

First we propose some simple static plots to observe the correlation patterns. In the following part, some interactive plot will allow a better investigation.

6.1.1 BRAF inhibition

Here is the version for publication:

And here is the version with table to replace interactive plot in the static publication:

6.1.2 Other targets

Additional plot for p53 and PI3K:

6.2 Interactive plots

We can have a deeper look at scatter plot with interactive settings

Here is the non-interactive reference plots

Let’s generate each column as an interactive plot, first with drugs and then with CRISPR:

And a last interactive plot to visualize the benefit of RNA addition for CRISPR prediction:

## [1] "Done"
## [1] "Computation time:"
## 143.629 sec elapsed